For most of its history, the Internet of Things has been a listening technology. Sensors measured temperature, vibration, location or flow. Networks carried those readings to a platform, and a person looked at a dashboard and decided what to do. Even the most advanced industrial deployments mostly worked this way: machines reported, humans acted.
That division of labour is starting to break down. A run of announcements in September 2026 points to the same shift. Connected systems are no longer only describing the physical world. Increasingly, they are expected to do something in it: move a part, steer a vehicle, stop a line, send a drone to check a pipe.
The industry has settled on a name for this: physical AI. It is a loose term, and it attracts its share of marketing. But underneath it sits a real change in what IoT infrastructure is for, and that change has consequences for everyone who designs, buys or operates connected systems.
What Physical AI Actually Means
Physical AI describes AI systems that perceive the physical world and act on it directly. Robots, autonomous vehicles, drones and automated machines are the obvious examples. What sets them apart from earlier automation is that their behaviour is not fully scripted in advance. They use models trained on large amounts of data to interpret what they see and choose what to do next.
IoT is the nervous system these machines depend on. A robot arm needs cameras and force sensors. An autonomous truck in a mine needs positioning, a network connection and data about the site around it. A warehouse drone needs to know where the shelves are and what is on them. Without the sensing, connectivity and data platforms that IoT built over the past decade, physical AI would have nothing to reason about.
The difference now is direction. Traditional IoT data flowed up, from devices to people. In physical AI systems, decisions increasingly flow back down, from models to machines, often without a person approving each step.
The Numbers Behind The Shift
The clearest signal this month came from Gartner. On 15 September, the research firm published its strategic predictions for 2027 and beyond, and physical AI featured prominently. IoT Business News reported on Gartner’s predictions for 2027 in detail.
Eighty Percent Of Front-Line Workers
Gartner expects that by 2030, 80% of front-line workers employed by international companies will be assisted by physical AI systems, including robots, drones and autonomous vehicles. Front-line work covers the jobs that happen on factory floors, in warehouses, on construction sites and in the field, rather than at a desk.
Predictions of this kind are forecasts, not measurements, and analyst timelines often slip. But the direction matters. If even a fraction of that figure materialises, the number of machines acting on IoT data will grow far faster than the number of dashboards reading it.
Ten Billion Autonomous Agents
The second figure is larger and more abstract. Gartner expects individuals, businesses and governments to create more than 10 billion autonomous agents by the end of the decade. Not all of these will touch physical equipment. Many will be software agents working in business systems. But a growing share will sit between data and machines, deciding when to reorder parts, reroute a vehicle or adjust a process.
Gartner’s framing for IoT is direct: connected infrastructure is moving from collecting telemetry to serving as an execution layer for AI-driven operations. That is a very different job description from the one most IoT platforms were built for.
Why The Edge Matters Now
A robot cannot wait for a round trip to a distant data centre before deciding whether to grip or release. Physical AI pushes computing closer to the machine, onto the device itself or onto a nearby gateway. Three announcements this month show how quickly that edge is maturing.
Agents On A Single Box
On 17 September, the MLPerf Inference v6.1 benchmark results added a new test for agentic workloads running at the edge. In the new test, NVIDIA’s Jetson AGX Thor developer kit ran a 27-billion-parameter language model through 1,007 turns of software-engineering tasks, as reported by IoT Tech News.
The full run took 24 minutes and 36 seconds using NVIDIA’s optimised software. The same workload on a standard open-source baseline took 2 hours and 37 minutes, a reduction of about 6.4 times. Accuracy on the function-calling tests, which check whether the model chose the right actions, came in at 87.94%.
Benchmark results always need care. They are run under controlled conditions, often on hardware configured for the test. Still, the fact that MLPerf added an edge agentic category at all says something. Miro Hodak, one of the benchmark’s inference co-chairs, put it plainly: “Complex inference systems with agentic properties are increasingly hosted on edge computing devices.”
Factories That Retrain Less
Speed is only one part of the problem. Factories change constantly: new products, new materials, new suppliers. An AI model that needs weeks of retraining every time a production line changes is of limited use.
At its AI Talk Concert in Seoul on 15 September, LG AI Research presented two manufacturing models built to cope with that. EXAONE Tabular analyses structured production and quality data, and EXAONE Omni-Inspect handles visual defect inspection. Both are designed to adapt to new conditions without full retraining.
EXAONE Tabular is small by current standards, at roughly 21 million parameters. LG says it achieves results similar to Google’s 1.64-billion-parameter TabFM model at about one-eleventh of the inference cost. Lim Woo-hyung, co-head of LG AI Research, framed the goal as solving “difficult problems that industries have struggled with for years.”
Smaller, more adaptable models matter for IoT because they can run closer to the machines producing the data. We explored how factory data turns into decisions in our profile of Datanomix and manufacturing data. Physical AI takes the next step: acting on those decisions directly.
Robots That Learn From One Video
The third development came from robotics. On 11 September, Skild AI launched S1, a robot foundation model trained on NVIDIA infrastructure. According to IoT Tech News, S1 can learn a new task from a single video demonstration, without retraining the model’s weights.
In multistep evaluations, the company reported an average per-step success rate of about 66%, against 9% for its baseline system. Skild says it reached an annual revenue run rate of $100 million within ten months of its commercial launch and has more than 60 deployment partnerships, including work with Foxconn on assembly tasks. Chief executive Deepak Pathak summed up the company’s view: “Learning by experience, and not preprogramming, is the step change that has happened in robotics.”
A 66% per-step success rate also shows how far there is to go. Across a long sequence of steps, those per-step failures compound. That is one reason the most serious physical AI deployments keep humans and hard safety limits in the loop, a point we return to below.
What Changes For IoT Networks
Physical AI does not only need more computing. It places new demands on the networks connecting machines, sites and platforms.
Private 5G Moves From Push To Pull
For years, private 5G was a technology looking for customers. Telecom vendors promoted it, and many enterprises concluded that Wi-Fi was good enough. RCR Wireless argues private 5G is now shifting from that “push” model to a “pull” model, driven by companies deploying robots and autonomous vehicles.
The article cites SNS Telecom & Research, which notes that robot vendors are now recommending private 5G themselves. It also lists deployments that fit the pattern: - Hiroshima Gas in Japan uses 5G-connected patrol robots to detect gas leaks. - Lyon-Saint Exupéry Airport in France uses valet parking robots. - China Huaneng Group runs 100 autonomous electric mining trucks.
The reasoning is practical. Mobile robots move between areas, need predictable latency and share spectrum with many other machines. Those are the conditions private cellular networks were designed for. We compared the options in our explainer on industrial IoT and industrial connectivity, and physical AI strengthens the case for managed, predictable networks on busy sites.
Low-power networks still have a role. Most sensors around a robot do not need 5G. As we covered in our LPWAN explainer on the networks built to whisper, the majority of IoT devices send small amounts of data infrequently. Physical AI adds a new, demanding layer on top of that base; it does not replace it.
Data That Crosses Borders
A less discussed problem is jurisdiction. Connected vehicles, drones and mobile robots move, sometimes across national borders, and the data they generate travels with them through foreign networks.
Writing in RCR Wireless in August, Curtis Govan, president of the Americas at connectivity provider floLIVE, described this as a gap between data sovereignty and network sovereignty. “Your data is sovereign at rest – and then the car drives across a border,” he wrote. His argument is that storing data in the right country is not enough if the data passes through networks and roaming identities in another jurisdiction on the way.
For physical AI, this matters more than for traditional IoT. A machine that acts on data needs that data to be timely and trustworthy, and regulators increasingly want to know where it flows.
The Hard Part: Safety And Control
The most important questions about physical AI are not about speed or scale. They are about what happens when the model is wrong.
AI Proposes, Embedded Software Decides
AI models are probabilistic. Given the same input twice, they may not produce exactly the same output, and they can be confidently mistaken. Traditional embedded control systems work the other way: they follow fixed rules and behave predictably.
A recent piece on what embedded software must solve for physical AI summarised the emerging design principle in four words: “AI proposes, embedded software decides.” In practice, that means an AI vision system might suggest a path for a mobile robot, but a simple proximity sensor wired to deterministic control logic can still override it and stop the machine.
This layered approach is not new in industrial safety. What is new is how much of the decision-making now sits in the probabilistic layer, and how important it becomes to keep the deterministic safeguards independent from it.
Testing Systems With No Single Right Answer
Testing is the other difficulty. Conventional software testing checks that a given input produces the expected output. AI systems often have no single correct output for every input, so teams must measure accuracy, false positives and negatives, robustness, latency and resource use together.
Validation also does not end at launch. Physical conditions change: lighting shifts, parts wear, new products arrive. Models that performed well in testing can drift once deployed. That makes continuous monitoring, collecting telemetry about how models behave in the field, a core IoT job in physical AI systems rather than an optional extra.
Accountability For Machine Actions
When a sensor reports a wrong reading, the damage is usually limited to a bad chart. When a machine acts on a wrong decision, the consequences are physical. That changes who carries responsibility.
Gartner’s predictions point in this direction. The firm expects that 80% of Global 500 companies will contractually assign their chief information officer or chief AI officer as an “evidence custodian” responsible for AI accountability. It also expects insurers to become important forces in shaping AI governance through the way they underwrite liability.
For IoT teams, the practical implication is that connected systems will need to record not just what devices measured, but what automated systems decided and why. Gartner describes a need for new identity, permission and audit capabilities for machine-generated actors. In plain terms, every agent and machine that can take an action may need its own identity, clear limits on what it is allowed to do and a reliable log of what it did.
Much of that plumbing already exists in some form in IoT: device identity, certificates, access control, event logs. The challenge is extending it from devices that report to systems that act.
What IoT Teams Should Watch
Physical AI is not arriving everywhere at once. Most connected devices will keep doing what they do today: measuring, reporting and waiting. But for organisations running robots, autonomous vehicles or automated sites, several questions are worth tracking over the next year.
The first is where decisions are made. Moving inference to the edge reduces latency and bandwidth, but it also spreads models across many devices that need updating, monitoring and securing.
The second is network design. Sites with mobile machines may need more predictable connectivity than Wi-Fi can offer, while the wider sensor estate can stay on low-power networks.
The third is governance. Teams should be clear about which decisions machines can take on their own, which need human approval and how every automated action is recorded.
The fourth is evidence. Vendor benchmarks and success rates are useful signals, but they are not the same as performance on a specific site, with specific equipment and specific people. Pilots that measure real outcomes remain the most reliable guide.
From Sensing To Acting
IoT spent more than a decade teaching the physical world to report on itself. Physical AI asks connected systems to go one step further and act on what they learn.
The announcements of the past few weeks show the pieces coming together: edge hardware capable of running agentic models, factory AI that adapts without constant retraining, robots that learn from demonstration and networks being rebuilt around moving machines. They also show the gaps, from success rates that still fall short of reliable operation to unanswered questions about safety, data and accountability.
For people who build, buy and invest in connected technology, the shift is worth watching closely. The value of IoT has always depended on what happens after the data arrives. Increasingly, the answer is that a machine does something with it.
